Efficacy Analysis in Different Architectures of Convolutional Neural Networks for COVID-19 Diagnosis Based on X-Ray Images

The Covid-19 pandemic has caused numerous infections and deaths worldwide since 2019 due to its high transmission and respiratory contagion capacity, reaching over 535.1 million confirmed cases. Nowadays, there are different types of tests to diagnose Covid-19, such as PCR or antigen tests. However, radiographs are characterized by being fast, efficient, and low-cost, being a key factor in the diagnosis of Covid-19. Deep learning can be added to this through convolutional neural networks that use convolution operations to classify data. This could mean a great advantage when diagnosing people who have the disease compared to healthy people, in order to support medical specialists and not overload the healthcare system. This article analyzes the efficacy of image classification through deep learning of three convolutional neural network architectures: standard, VGGI6, and NASNet. The database was collected from different repositories such as GitHub, Radiopaedia, the Italian Society of Radiology (SIRM), and others, gathering a total of 10,000 X-ray images between healthy patients and patients with Covid-19. The Matlab software was used for image preprocessing and the Google Colaboratory web application for deep learning training. The proposed convolutional neural networks achieved an accuracy of 97%, 94.5%, and 93.3%, respectively, which shows high effectiveness in classifying X-ray images with Covid-19 from healthy patients.